The database searches yielded 1024 articles as shown in Fig. 1. After removal of 27 duplicates, 972 articles were excluded on the basis of not fulfilling inclusion criteria, resulting in 25 full-text articles for inclusion. From the reference lists of these articles, 5 additional articles were identified for inclusion, resulting in a total of 30 articles for analysis. As shown in Fig. 2, the majority of included records (N = 23) were published in the last 4 y (2016–2019).
Of the 30 included publications, the majority were prospective studies (n = 24) while the remaining six studies were retrospective. Approximately half employed a randomized control group, of which 13 were randomized controlled clinical trials and three randomized vignette-based studies, where clinicians were presented with randomly assigned hypothetical scenarios. One of the randomized controlled trials was undertaken in a simulated environment .
Five studies employed a control group without randomization, of which four were pre-post studies that observed a control cohort during the same period as the intervention cohort. The remaining nine articles were pre-post studies without a control cohort (see Table 1). Three of the pre-post studies employed a time-series analysis, comparing temporal trends as well as outcome differences before and after the intervention [30,31,32].
Summary of included studies
Table 2 summarises the characteristics of the included studies. The complete data set is available in supplementary materials. Half of the studies were set across hospitals and tertiary care centres and targeted a diverse population of HCPs (physicians, nurses, midwives, medical assistants, novice health care providers, entire clinical team). Eleven studies were set in the context of ambulatory care. One study covered both primary care and specialty clinics, and one study covered an entire ambulatory healthcare system. The remaining four studies were conducted outside clinical environments: one study recruited novice healthcare providers affiliated with a medical school to conduct an RCT in a simulated setting , one study recruited practitioners from a collaborating hospital to undertake a vignette-based study , and two studies recruited practitioners across geographical regions, namely the greater Chicago area  and across Australia  to undertake vignette-based studies.
Sixteen of the studies delivered their intervention via Electronic Health Record (EHR) systems, either through the prescription and ordering interface or through patient records. Six studies delivered the intervention mainly through modifications in the physical environment such as posters [36, 37], aromatization [29, 38, 39], or props . Four studies combined other strategies with email communication in order to provide feedback and statistics on the performance of the target behaviour [37, 41,42,43]. One study utilized email messages to direct participants to an electronic dashboard . One study delivered its intervention solely through an email message . And one additional study delivered its intervention solely through a letter .
The two studies targeting modification of ventilation practices included the modification of default setting on the ventilation machines [30, 42] One additional study delivered a cost awareness message printed on test results . Two studies delivered their interventions mainly through the modification of routine activities, namely the modification of syringe sizes for sedation during endoscopy procedures , and modification of inpatients’ printed charts with pain management procedures .
Nudging strategies and objectives
In total, 43 nudges were used in the 30 included studies, since 10 studies employed two or three nudging techniques each. From these we identified and coded 11 different nudging strategies, described in Fig. 3 and Table 3. The large majority of studies employed only one nudging strategy (n = 20). From 2016 onward, 7 studies used 2 different nudging strategies, and 3 studies included 3 nudging strategies each (See Table 2). All but one of these” multi-nudging” studies had a positive result. Of all nudges, 84% resulted in statistically positive results.
Since outcomes were very heterogeneous, we clustered articles into four different objectives to facilitate comparison. The majority of studies had as their objective to change prescription and ordering behaviour, namely, encouraging judicious antibiotic prescription [34, 36, 37, 41], increasing vaccinations orders [49, 50], increasing prescription of generic medication , reducing prescription of high-cost, low-value medication [32, 52], reducing unnecessary laboratory tests [33, 53], reducing imaging procedures [31, 54], increasing prescription of mouthwash to intubated patients in the intensive care unit (ICU) , increasing guideline-concordant prescription of statins , increasing prescription of blood glucose (A1C) tests for diabetes prevention , and increasing high-value treatment for lower-back pain .
The second largest target was the modification of behaviours with respect to certain care procedures (n = 7) such as ventilation settings for intubated patients [30, 42], improving inpatient sleep , sedation during endoscopy , screening for risk of cardio-vascular disease in primary care , sitting down during examinations , and pain management after Caesarean Section surgery . Four studies targeted hand hygiene [29, 38, 39, 43], and two studies focused on vaccination of healthcare providers [45, 46].
Of the 17 studies with the objective of changing prescription and ordering behaviour all but two had a successful outcome (statistically significant positive results). Of the 7 studies targeting the modification of behaviours with respect to certain care procedures, only one had a non-significant result. Of the 4 studies targeting hand hygiene, three were successful. Of the 2 studies targeting vaccination of healthcare providers none had a positive result.
During the analysis, we found that some strategies appealing to the analytical System 2 could only cause impact if the participant chose to pay attention. For example, only showing the cost of a particular medication on the ordering screen (information transparency) can go completely unnoticed by a participant thereby not engaging the participant’s analytical mind. We also found that some strategies that automatically substituted brand medications for generics or automatically populated ordering forms (defaults/pre-orders) could also take place without the participant’s knowledge or awareness. Thus, these interventions could change the outcome of a certain behaviour without having any effect on the behaviour itself.
Another challenge with considering System 1 and System 2 categories is that they do not capture whether the intervention is presented at the time of the decision or not. For example, some interventions sent feedback emails that could be read anytime, not at the time of prescription. So, the emails are meant to change a belief that will eventually impact a behaviour, but that is very different from immediately changing a prescription to generic at the time of ordering. To capture these shortcomings, we instead developed two independent dimensions, creating four quadrants:
Synchronous vs. asynchronous - An intervention strategy was coded as synchronous if its delivery coincided with the decision or behaviour it intended to affect.
Active vs. passive - An active strategy cannot be completed without an action on the part of the participant.
Figure 4 shows the distribution of nudging objectives across the quadrants. With the exception of staff vaccination, the other nudging objectives were addressed by strategies in more than one quadrant.
Figure 5 shows the distribution of intervention strategies across the quadrants. Three intervention strategies were assigned to more than one quadrant: alerts/reminders, defaults/pre-orders, environmental cueing/priming. Alerts and reminders often take place when the target behaviour must be performed (synchronous) and require an action from the clinician (active). Two exceptions were observed. One study was coded as active and asynchronous (Q2)  as it sent an email reminder prompting clinicians to access a dashboard in order to be exposed to the other nudges. The other study  was coded as passive and asynchronous (Q3), where letters were sent to remind front line staffers to get vaccinated, the letters could be read at any time and no action was strictly required.
Most of the environmental cueing/priming strategies are passive and asynchronous (Q3), for example, placing posters with crafted messages and pictures. However, we coded one study  as synchronous because it placed chairs in examination rooms to encourage participants to sit during examination (Q4). In this case, the priming object (chair in examination room) coincides in time with the target behaviour (sitting during examination). The majority of default/pre-order interventions were synchronous and passive (Q4), for example, automatically changing a brand name medication for its generic equivalent. The” nudge” (substituting for generic) coincided in time with the target behaviour (prescribing a medication), and happened automatically even if there was no action from the clinician. There were two exceptions in this category, which implemented asynchronous strategies. One study  periodically populated an electronic folder with pre-approved orders for A1C laboratory tests, the other study  sent letters to participants with either pre-booked appointments (opt-out) for vaccination or without pre-booked appointments (opt-in).